NVIDIA Releases SANA: 20× Smaller, 100× Faster Than Flux-12B
NVIDIA has open-sourced the SANA model family across four variants. SANA-Sprint produces 1024px images in 0.1 seconds on H100 (0.3s on RTX 4090) and runs in under 8GB VRAM at 4-bit quantization. SANA-Video outperforms 14B-parameter competitors with a 2B model at 36s versus 1,897s latency. SANA-WM generates 720p one-minute video with camera control. Sol-RL post-training delivers 4.64× faster convergence. Day-one integration with Diffusers, ComfyUI, and SGLang.
Why It Matters
SANA dramatically lowers the hardware bar for production-quality image and video generation. Running competitive image generation on a consumer RTX 4090 in a fraction of a second opens the stack to local inference deployments previously impractical at this quality level.